Building Interpretable AI Systems

Despite the popularity and recent advances in Artificial Intelligence (AI) systems boosted by machine learning (ML) methods, most of the existing models fall short on their ability to explain their reasoning process, i.e., in providing human-like justifications for the reasoning behind a certain AI task. This functionality of "being interpretable" is a fundamental requirement for the adoption and uptake of AI systems in real-world scenarios, as users need to trust and understand the approximations and inferences done by the system.

The application of AI in contexts with high social and economic impact (as in health care and legal settings) will require the evolution of black-box AI models in the direction of systems which can justify, explain and dialogue with their end-users about the underlying reasoning process, providing transparent human-interpretable outputs.

This project aims at designing, building and evaluating a framework for the construction of Interpretable AI systems, with an emphasis on Natural Language Processing (NLP) tasks. The goal is to support the construction of complex AI systems for addressing tasks such as Question Answering and Text Entailment, which can output meaningful human-like explanations in addition to the expected output. Part of the project will involve the investigation of interpretable machine learning approaches for NLP, e.g., through exploratory analysis and visualisation of intermediate results returned by a ML model and improvement of model architecture design.